61 results about "Row vector" patented technology

The invention provides a water supply network water leakage abnormity identification method based on NMF dimension reduction, and relates to the technical field of water supply network abnormity identification, and the method comprises the steps: embedding hydrophone sample data containing expert labeling information into a sample weight matrix as a row vector, and constructing a sample dimension loss item; hydrophone sample data containing industry feature priori are mapped to the feature basis matrix to serve as rows, and feature dimension loss items are constructed; combining the two loss items to construct a total loss function and solving the total loss function to obtain an optimal sample weight matrix; and performing clustering analysis based on the optimal sample weight matrix to obtain a clustering affiliation result of the hydrophone sample data so as to judge whether the water supply network to be identified has water leakage abnormity or not. According to the method, the double-constraint NMF framework fusing the expert knowledge of the sample dimension and the industry priori knowledge of the feature dimension is constructed, the accuracy and interpretability of hydrophone feature dimension reduction are considered, and the problems that an existing method is poor in interpretability, less in priori knowledge utilization and the like are solved.

The invention belongs to the technical field of defect detection, and particularly relates to a defect detection method and system in a filter assembly laminating process based on image processing, so as to solve the technical problems that an existing detection method model is relatively weak in micro defect sensing capability and relatively low in detection accuracy. The detection method comprises the following steps: S1, carrying out block attention calculation on an image block sequence; s2, calculating an adjustment coefficient and adjusting the updating process of the row normalization factor by using the adjustment coefficient; s3, weighting corresponding row vectors in the initial context feature matrix by using the adjustment coefficient corresponding to each query block to obtain a weighted context feature matrix; and S4, outputting a mask pattern representing the position of the defect on the surface of the filter assembly. According to the detection method provided by the invention, secondary weighting and feature fusion are carried out on the initial context features, so that the accuracy of defect detection of the filter assembly is improved.

The application discloses a new correlation image recognition and machine learning method based on fuzzy monotony, and belongs to the technical field of artificial intelligence of pattern recognition and machine learning; the method is defined as FMMCA, which evaluates local fuzzy monotone correlation by comparing row vectors and column vectors of an image matrix pair by pair, then the local correlations are weighted and accumulated, and finally the global fuzzy monotone correlation between images is obtained. The application directly uses fuzzy monotone correlation analysis to replace classical correlation analysis for multi-view research, so that the problems existing in classical correlation analysis do not exist, and the fuzzy monotone method feature does not need to be measured statically by distance, but can be measured dynamically by interval change, so that the influence of some noise is offset, the performance is improved, a new fuzzy monotone machine learning method is formed, the optimization of a focus loss function is not needed, the parameters are few, the robustness is good, and the computing power is small.

The invention relates to the technical field of electrical digital data processing, and discloses a Web application automatic generation and data processing method based on spreadsheet mapping, which comprises the following steps: mapping spreadsheet cells into enumeration values to construct a type fingerprint matrix; calculating a type purity index of a column vector, identifying a vertical blocking band, and splitting the matrix into a plurality of orthogonal sub-matrixes which are not communicated with each other; calculating a type distribution difference value of adjacent row vectors in the sub-matrix to determine a horizontal logic partition line so as to define a logic data block; a Web application object model is assembled based on an arrangement mode of logic data blocks, signal crosstalk between transverse side-by-side data entities is cut off by introducing a vertical topology decoupling mechanism based on column entropy, and the technical problem that downlink feature scanning is easily interfered by orthogonal noise in a multi-table mixed arrangement scene is solved. And the deterministic structure reconstruction of the complex two-dimensional heterogeneous data under the condition of no preset template is realized.

The invention provides a bearing fault identification method based on singular value decomposition. The method comprises the following steps: acquiring a discrete vibration signal of a bearing; a Hankel matrix of the vibration signals is constructed, the dimension of the matrix is smaller than < and is the length of the vibration signals, and the column number is determined based on the minimum characteristic frequency of typical components of the equipment; performing singular value decomposition on the matrix; selecting an effective row vector of the matrix by taking high periodicity, high cyclic stability and low complexity of the signal as targets, determining an effective singular value based on the effective row vector, and performing signal reconstruction based on the effective singular value; and carrying out spectrum analysis on the reconstructed signal and carrying out bearing fault identification according to a relation between a prominent frequency component in a spectrum and a bearing fault characteristic frequency. According to the bearing fault identification method, the features of the signals can be enhanced while noise reduction is carried out on the signals, and accurate judgment of bearing faults is facilitated.

The invention provides a large language model binarization quantification method and system based on instructive alternate optimization, and relates to the technical field of large language model deploying.The method comprises the steps that a weight matrix is divided into an important area and an unimportant area by evaluating the influence degree of all weight parameters on the performance of a large language model; carrying out binaryzation on all areas of the weight matrix, and alternately optimizing a row vector scaling factor and a column vector scaling factor by adopting a first-order row-column alternate optimization iteration mode to obtain a first-order reconstruction weight matrix; and carrying out binaryzation again on the important region of the first-order reconstruction weight matrix, carrying out optimization by adopting a first-order and second-order row-column alternating optimization iteration mode to obtain a second-order reconstruction weight matrix, and correspondingly taking the second-order reconstruction weight matrix as a weight parameter after the large language model is quantized. The quantization error of the key weight parameter can be effectively reduced, the quantization precision is high, the quantization process completely depends on the internal structure information of the large language model, and the edge deployment compatibility is high.

The application relates to the technical field of electric digital data processing, and discloses a Web application automatic generation and data processing method based on electronic table mapping, which comprises the following steps: mapping electronic table cells into enumeration values to construct a type fingerprint matrix; calculating a type purity index of a column vector and identifying a vertical blocking band, and splitting the matrix into a plurality of mutually unconnected orthogonal sub-matrices; calculating a type distribution difference value of adjacent row vectors in the sub-matrices to determine a horizontal logical segmentation line, so as to define a logical data block; and assembling a Web application object model based on an arrangement mode of the logical data block. The application introduces a vertical topology decoupling mechanism based on column entropy, cuts off signal crosstalk between horizontally arranged data entities, solves the technical problem that a horizontal feature scanning in a multi-table mixed arrangement scene is easily interfered by orthogonal noise, and realizes deterministic structure reconstruction of complex two-dimensional heterogeneous data under the condition of no preset template.

The invention provides a data intelligent analysis method and system applied to carbon emission monitoring, and the method comprises the following steps: executing an NMF load decoupling algorithm to obtain a dynamic non-negative vector which comprises a resident coefficient and an industrial coefficient; according to the non-negative row vector and the space-time database, respectively calculating resident and industrial carbon emission intensities corresponding to the current time period, and updating the carbon emission intensities to the space-time database; calculating and generating future carbon emission predicted values corresponding to different future time periods according to the space-time database; and performing anomaly detection calculation according to the carbon emission intensity and a future carbon emission predicted value corresponding to the current time period. By adopting the method, the carbon emission intensity condition of the whole area can be intuitively and indirectly reflected in real time without tampering.

A kind of GNSS data processing method based on single observation filtering, comprising the following steps: first, design matrix, observation vector and covariance matrix are decomposed into multiple row vectors or scalar, then one observation is processed each time, after all the observations of current epoch are processed through multiple cycles, the state vector and covariance matrix of current epoch recursive update are obtained;GNSS parameters in state vector are divided into common parameters and non-common parameters, when processing each observation, update common parameters, then gradually increase new parameters in state vector, i.e. process observation by observation, to obtain extended state vector and covariance matrix;In the third step, when data processing is carried out using the above-mentioned extended state vector and covariance matrix, it includes single epoch processing and multi-epoch processing;The state estimation of current epoch is output as the result of GNSS positioning.The present application has high efficiency when processing high-dimensional observation data.

The application provides a modulation method and system of an orthogonal frequency division multiplexing signal, and belongs to the technical field of wireless communication. The method comprises the following steps: mapping a frequency domain data signal and a pilot signal to be transmitted into a frequency domain resource grid matrix; performing a frequency domain shift operation on each row vector of the frequency domain resource grid matrix, shifting a zero frequency component to the center of the vector, and obtaining a shifted matrix; and performing an inverse fast Fourier transform on each row vector of the shifted matrix, and obtaining a time domain orthogonal frequency division multiplexing symbol matrix as a modulation result. The application constructs the entire frequency domain resource into a two-dimensional matrix, and all subsequent signal processing operations are based on the row vectors of the two-dimensional matrix as a basic unit, and are completed through a series of explicit matrix and vector operation steps. Thus, the IFFT call for the entire frame, which is a black box, is decomposed into individual steps for a single symbol, which are visible and controllable, and the problem of opaque process is solved.

The present application relates to the field of wastewater treatment, and more particularly to a method and system for online analysis of inkjet printing wastewater treatment data, the method comprising: obtaining water quality, process and equipment state parameter data of inkjet printing wastewater based on each sensor, and constructing an original data matrix after preprocessing; decomposing the matrix, selecting a preset principal component number for reconstruction, and generating a reconstructed data matrix; calculating reconstruction error by analyzing the difference between the original and reconstructed matrices, analyzing the correlation distortion index based on row vector similarity, constructing a target function to determine the optimal principal component number; and using reconstruction error and distance for clustering, filtering abnormal data points and adjusting process parameters. The present application effectively solves the noise interference problem under complex water quality changes by optimizing the principal component number and improving the clustering analysis, significantly improves the abnormal detection accuracy and system stability, and enhances the robustness and adaptability of the inkjet printing wastewater treatment system.

The invention provides a general interference suppression method and system based on adaptive window short-time Fourier transform, and the method comprises the steps: carrying out the adaptive initialization of a Gaussian window width parameter according to the amplitude standard deviation of each pulse signal, carrying out the short-time Fourier transform, obtaining a time-frequency matrix of each pulse, carrying out the vectorization, carrying out the arrangement according to columns, and obtaining a time-frequency matrix of each pulse; constructing a time frequency-pulse matrix; performing complex RPCA decomposition on the matrix, and separating a low-rank matrix and a sparse matrix; adjusting a Gaussian window width parameter according to energy distribution feedback of each column vector of the sparse matrix, and performing iterative optimization until convergence to obtain a final sparse matrix; extracting interference time-frequency column vectors, reconstructing the interference time-frequency column vectors into an interference time-frequency matrix, and performing inverse short-time Fourier transform to obtain a time-domain interference signal; and finally, subtracting the corresponding interference signal from the original echo signal. According to the method, various types of interferences can be effectively suppressed, and particularly, the suppression effect can be remarkably improved on the premise of protecting target signals when broadband interferences and complex deception interferences are handled.

A kind of matrix computing device, method, system, circuit, chip and equipment, compressed format matrix can be directly calculated, so as to improve the computing efficiency of compressed format matrix.Matrix computing device includes: vector outer product processing engine and accumulator, matrix computing device is based on vector outer product and carries out the calculation of compressed format first matrix and second matrix, in the process of calculation, vector outer product processing engine carries out vector outer product calculation to the first column vector of the row coordinate reserved and the second column vector of the column coordinate reserved, then accumulator is based on the index of position coordinate, and the third element value of the same position coordinate is accumulated, to obtain the result matrix of two compressed format matrixes for calculating, relative to the method that compressed format matrix needs to be decompressed first in traditional, then matrix calculation is carried out to the decompressed matrix, the matrix computing device provided in the embodiment of the application can effectively improve the computing efficiency of compressed format matrix.

The application discloses a text object recognition method and device, which is used to improve the efficiency and accuracy of type identification of objects included in network attack events or dynamic events. The method comprises the following steps: in response to a recognition instruction, extracting a plurality of objects included in the description text of a network event; respectively extracting features of the plurality of objects to determine a feature matrix of each object; wherein each row vector in the feature matrix of any object is used to represent a feature of any object; performing feature re-extraction on each column vector of the feature matrix of any object to generate an intermediate matrix corresponding to each column vector, and performing feature enhancement on each intermediate matrix to generate a target vector corresponding to each intermediate matrix; inputting a target matrix composed of the generated plurality of target vectors into a pre-trained type prediction model to determine the type to which any object belongs.

The application discloses an integrated diffraction neural network and processor for matrix calculation, the diffraction neural network comprising: an input module configured to encode row vectors or column vectors of an input matrix into input optical signals, the input optical signals encoded by different row vectors or column vectors having different wavelengths; the diffractive optical transformation module is configured to transform optical mapping of the operator matrix, and the diffractive optical transformation module is used for receiving input optical signals, applying the same optical transformation to the input optical signals with different wavelengths and converting the input optical signals into output optical signals; and the output module is used for receiving and separating the output optical signals with different wavelengths and guiding the output optical signals to a detector for detecting the signals. As the diffraction type optical conversion module applies the same optical conversion to the input optical signals with different wavelengths, conversion operator matrixes are consistent for the input optical signals with different wavelengths within a certain bandwidth, and the optical conversion efficiency is improved. And the input signal light with different wavelengths can obtain a result of executing corresponding matrix calculation of the input matrix and the transformation operator matrix through the modulation unit. The input module, the diffraction type optical transformation module and the output module in the integrated diffraction neural network for matrix calculation are easy to integrate, and the minimum integration area of the integrated diffraction neural network for matrix calculation can be reduced. The matrix calculation process is a passive process, and the system energy consumption of large-scale matrix calculation is remarkably reduced.

The invention discloses a hypergraph low-rank sparse feature selection method based on a kernel function, and the method comprises the steps: converting an attribute set of a data set X into a kernel matrix K through kernel mapping, and obtaining an attribute self-expression matrix Q according to attribute self-expression; obtaining a normalized feature similarity matrix Lnorm by the hypergraph modeling data structure; splitting the attribute self-expression matrix Q according to the low-rank constraint, namely Q = AB, and solving an optimization matrix AB; sparse selection is carried out on all row vectors of the matrix AB, and when the absolute value sum of all elements of a certain row vector is smaller than the 0.1-time mean value of the absolute value sum of all elements of the matrix AB, it is considered that the data association capacity of the row elements is poor, and the row elements can be removed; the obtained matrix AB is an optimal feature subset obtained by screening, SVM classification is carried out by using the feature subset, and the classification accuracy is evaluated. According to the method, the feature subset representation capability can be improved, feature support is provided for subsequent data mining and machine learning tasks, and a foundation is laid for effective analysis and application of high-dimensional data.

The application discloses a data processing method and device, electronic equipment and a storage medium, and relates to the technical field of computers. A specific implementation manner of the method comprises the following steps: in response to a data processing request, obtaining a corresponding to-be-processed array, and generating a matrix corresponding to each to-be-processed array; wherein each to-be-processed array comprises at least two data, each data in the to-be-processed array is in one-to-one correspondence with a row vector in the corresponding matrix, and the arrangement order of the row vectors in the matrix is the arrangement order of the corresponding data in the to-be-processed array; calculating the transpose matrix of each matrix, performing a preset Boolean operation on the row vectors at the same positions in the transpose matrices in sequence, and obtaining an operation result matrix; transposing the operation result matrix, obtaining a data processing result, and sending the data processing result. The implementation manner can solve the problem that when the amount of data that needs to be operated is large, the operation mode of a single bit will take a long operation time and the efficiency is low.

The invention discloses a navigation signal compression and capture method based on a local pseudo code measurement matrix, and the method comprises the steps: firstly carrying out the cyclic shift of a navigation signal and a pseudo code, and respectively constructing a navigation signal sampling point matrix and a code phase matrix; and sparse transformation is carried out on the navigation signal through matrix multiplication between the navigation signal sampling point matrix and the code phase matrix to obtain a correlation matrix. Then, a greedy strategy is used for selecting M rows of vectors with the maximum code matrix and the lowest correlation coefficient to construct a measurement matrix, and compression measurement is carried out on the correlation matrix through the measurement matrix constructed through the greedy strategy; secondly, reconstructing the correlation matrix by using a subspace tracking algorithm and a least square operator; and finally, selecting the ratio of the primary peak value to the secondary peak value as a threshold value, and when the ratio of the primary peak value to the secondary peak value of the reconstructed correlation matrix is greater than the threshold value, outputting Doppler frequency offset and code phase offset. Otherwise, the navigation signals are captured again. According to the method, the maximum correlation coefficient of the measurement matrix can be reduced, the navigation signal capturing sensitivity is improved, and hardware resources in the signal capturing process in the navigation receiver can be reduced.

The application provides a power system abnormal data recovery method and system considering time and precision priority, and belongs to the technical field of power data security. The method comprises the following steps: obtaining power data to determine whether there is an abnormal value; obtaining the quantized value of each abnormal value in two dimensions based on a security protection level index and a response recovery level index to form a row vector, and further establishing a decision index matrix; determining a weight matrix by using an analytic hierarchy process; calculating the comprehensive score of each abnormal value based on the decision index matrix and the weight matrix; determining a priority label according to the comprehensive score and a threshold value, performing label processing on each abnormal value in the power data set, and obtaining a data set with a priority label; recovering each abnormal value in the data set by using a recovery algorithm corresponding to the priority label to obtain a normal data set; the recovery algorithm comprises a high-precision recovery algorithm and a fast recovery algorithm; and the normal data set is sent to a data control center to provide a data source for subsequent real-time decision analysis of the power system.

The application provides an eyelid tumor classification method and system, relates to the technical field of medical image analysis, and comprises the following steps: feature map blocking; main similarity matrix construction; dense feature similarity matrix construction; cosine angle calculation between the main similarity matrix and the corresponding row vectors of the dense feature similarity matrix in units of behaviors; definition of a partial cosine adjustment formula according to the cosine angle to globally fine-tune each element in each row vector of the dense feature similarity matrix; after adjustment of each element in units of behaviors, an adjusted dense feature similarity matrix is obtained; similarity calculation of each category feature of the query set and the support set according to the adjusted dense feature similarity matrix to obtain a classification result, and output of the probability of eyelid tumor belonging to different categories through a Softmax function, so that the query set image is determined as the category with the maximum probability. The accuracy and reliability of eyelid tumor benignity and malignancy classification under small sample conditions are greatly improved.

The invention provides a visible light sight distance signal identification method based on machine learning and a related device. Receiving an optical signal carrying a synchronous training symbol; deriving a received signal strength from the received optical signal; the method comprises the following steps: performing synchronization operation on received optical signals to obtain a timing metric value, intercepting through a sliding window according to a peak value of the timing metric value to obtain a timing metric peak value sequence, and extracting characteristic quantities of the timing metric peak value sequence, including a peak value, a mean value, a standard deviation, kurtosis, skewness and a median; fusing the received signal strength and the characteristic quantity of the timing metric peak sequence to obtain a fused row vector; performing normalization processing on the fused row vector to obtain a mixed feature vector; inputting the mixed feature vector into a trained recognition model, and performing visible light sight distance signal recognition according to the mixed feature vector to obtain a recognition result; the average accuracy of signal recognition is improved, and the problem that the recognition success rate of a sight distance signal at the edge in a complex optical environment through a single RSS feature is low is effectively solved.

The invention discloses an image redirection method based on graph saliency sorting and reinforcement learning. Firstly, a visual shape graph and a visual semantic graph are constructed with instances as nodes, the two graphs share an adjacent matrix, node features of the visual shape graph are instance area features, and node features of the visual semantic graph are instance visual semantic fusion features; then, performing weighted summation on the visual shape graph node embedding matrix, the visual semantic graph node embedding matrix and the shared node embedding matrix to obtain a final node embedding matrix; finally, each row vector of the node embedding matrix is an instance node embedding, and each instance node embedding passes through a full connection layer to obtain a significance score of each instance; and finally, realizing image redirection based on reinforcement learning. Through cooperation of graph structure saliency sorting and saliency guide reinforcement learning, unified improvement of structure, semantics and visual beauty is realized, and the problems of rough saliency information, unreasonable operator selection, distortion of redirection results and the like are solved.

The invention provides an eyelid tumor classification method and system, and relates to the technical field of medical image analysis. Constructing a main similarity matrix; constructing a dense feature similarity matrix; calculating a cosine angle between corresponding row vectors of the main similarity matrix and the dense feature similarity matrix by taking a row as a unit, defining a partial cosine adjustment formula according to the cosine angle so as to perform overall fine adjustment on each element in each row vector of the dense feature similarity matrix, and adjusting each element by taking a row as a unit; obtaining an adjusted dense feature similarity matrix; and according to the adjusted dense feature similarity matrix, calculating the similarity of each category feature of the query set and the support set to obtain a classification result, outputting the probability that the eyelid tumor belongs to different categories through a Softmax function, and judging the query set image as the category with the maximum probability. The accuracy and reliability of benign and malignant eye tumor classification under the small sample condition are greatly improved.

A kind of data service recommendation result diversification method based on determinant point process, comprising the following steps: first, construct service recommendation score auxiliary matrix and service function similarity auxiliary matrix;Second, construct the service information kernel matrix required by determinant point process calculation, process determinant calculation result, optimize calculation process;Third, using cholesky decomposition, quickly judge in the process of greedy iteration, generate diversified reordering service recommendation result.The present application grasps the diversification demand under the service recommendation scene, as a kind of post-processing method, can carry out diversification processing on the basis of existing service recommendation model;While considering the functional correlation and diversity between services;Using the optimized solving method based on decomposition factor and row vector, the diversification process is optimized, and the calculation consumption of each iteration is reduced;For the existing service recommendation algorithm with recommendation accuracy as index, the recommendation supplement of diversity demand is carried out.

The invention discloses an ERT inversion image imaging method and system based on sensitivity optimization, and the method comprises the steps: classifying boundary measurement values, constructing an initial sensitivity matrix with a full air field as a reference, and carrying out the targeted zero setting and enhancement optimization of a column vector and a row vector according to liquid level information, according to the method, the problem of measurement failure caused by the fact that an electrode makes contact with a gas phase when a sewage pipeline is in a non-full-field state is solved, the problem of inapplicability of the ERT technology under the non-full-field working condition is solved, the problem that the imaging speed and the imaging precision cannot be achieved at the same time is solved, a sensitivity self-adaptive adjustment algorithm is used, and the measurement accuracy is improved. In combination with an image reconstruction algorithm of one-step iteration, the resolution of an inversion image can be effectively improved while the imaging speed is ensured, so that real-time and high-precision detection of a detected field domain is realized.

The application belongs to the technical field of finance, and particularly relates to a kind of announcement title generation method, device, terminal and medium.The announcement title generation method comprises: determining to be analyzed file in the file of announcement according to file type;Text content of the to-be-analyzed file is converted into a set of row vectors;According to font type and layout style, the row vectors in the set of row vectors are filtered to determine target row vectors;Generate announcement title based on the target row vectors.In this way, the application analyzes the text content of the announcement file by text analysis, to obtain key content in the announcement file as the announcement title, so as to ensure that the announcement title can accurately reflect the announcement content, and improve the convenience of users to obtain announcement content.

The invention relates to the technical field of ATE testing, in particular to a self-adaptive correction method and system for ATE test vectors and a storage medium. A test is automatically executed and a result is monitored (S1), when the test fails, invalid memory information is automatically acquired in a blocking manner, and correction content is analyzed and determined based on actual level data, so that a hard process that an engineer manually locates an error in millions of rows of vectors is replaced. And then a correction instruction is generated and directly sent to the ATE equipment for online correction, so that the links of exporting the vector file from the ATE equipment, modifying the vector file in an external tool and recompiling and loading the vector file for a long time in the traditional method are omitted. The series of automatic steps are combined with the control logic of iterative execution, so that the correction efficiency is improved.

The invention discloses a road surface green low-carbon evaluation method. The method comprises the following steps: S1, selecting road surface carbon emission evaluation indexes based on related specifications of road construction, operation and maintenance processes; s2, carrying out index classification on the selected road surface carbon emission evaluation indexes according to multiple dimensions; s3, determining the weight of each index; s4, constructing an index hierarchy analysis and evaluation model; s5, respectively establishing an evaluation factor set and a weight set according to the index structure, hierarchical division and index weights; s6, according to the difference between the characteristics of the evaluation object and the actual situation, setting the evaluation standard as a comment set with different grades; s7, evaluating the index layer according to the comment set, performing quantitative processing on the indexes to obtain the membership degree of the ith factor to the jth evaluation, and establishing a fuzzy relation matrix; s8, performing fuzzy operation on the weight and the fuzzy matrix to obtain a row vector index layer fuzzy evaluation result; and S9, determining a final evaluation result of each level index.

The invention belongs to the technical field of parameter identification, and provides a parameter identification method and system for a hybrid linear dynamic system, and the method comprises the steps: obtaining a fitting high-dimensional time sequence track of the hybrid linear dynamic system; constructing a hybrid linear dynamic system model based on the obtained fitting high-dimensional time sequence track to obtain a high-dimensional state transition matrix; disassembling a parameter estimation task based on a hierarchical idea and a row decomposition strategy, and disassembling the obtained high-dimensional state transition matrix into a low-dimensional row vector estimation task; updating parameters of the disassembled low-dimensional row vector by adopting a mobile data window increment stochastic gradient algorithm; a self-adaptive step length mechanism is introduced, the step length is dynamically adjusted and updated based on the parameter estimation residual error, and a prediction residual error is obtained; and estimating a transfer matrix and a noise covariance of the hybrid linear dynamic subsystem according to the obtained prediction residual error, and completing parameter identification of the hybrid linear dynamic system.

A systolic array and an accelerator including the same are disclosed. The systolic array may include n×n processing elements disposed in an n×n matrix (n being an integer equal to or more than at least 4), wherein the n×n processing elements perform a first convolution operation on first input data of row vectors of a first input n×n matrix and first weight data of column vectors of a first weight n×n matrix to generate n first output data, or each of at least k partial systolic arrays (k being an integer equal to or more than at least 4) constituted by dividing the n×n processing elements includes m×m processing elements disposed in an m×m matrix (m being an integer less than n and equal to or more than at least 2).